Journal of Process Control 12 (2002) 519531

www.elsevier.com/locate/jprocont

A decade of progress in iterative process control design:

from theory to practice
Michel Gevers*
CESAME, Universite Catholique de Louvain, Batiment EULER, 4 av. Georges Lemaitre, B-1348 Louvain-la-Neuve, Belgium

Abstract
One of the most active areas of research in the nineties has been the study of the interplay between system identication and
robust control design. It has led to the development of control-oriented identication design, the paradigm being that, since the
model is only a tool for the design of a controller, its accuracy (or its error distribution) must be tuned towards the control design
objective. This observation has led to the concept of iterative identication and control design and, subsequently, to model-free
iterative controller design, in which the controller parameters are iteratively tuned on the basis of successive experiments performed
on the real plant, leading to better and better closed-loop behaviour. These iterative methods have found immediate applications in
industry; they have also been applied to the optimal tuning of PID controllers. This paper presents the progress that has been
accomplished in iterative process control design over the last decade. It is illustrated with some applications in the chemical industry. # 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Identication for control; Validation; Iterative controller tuning

1. Introduction
Iterative process control is a control design methodology that has emerged in the nineties as a result of intense
research eorts aimed at bridging the gap between system
identication and robust control analysis and design. In
order to give the reader an idea about how wide this gap
was, we quote from a keynote lecture delivered at the 1991
IFAC Symposium on Identication [1].1
The last ten years have seen the emergence of
robust control theory as a major research activity.
During the same period, research in system identication has dwindled, and it might be tempting to
believe that most of the theoretical questions in
identication theory have been resolved for some
time. The surprising fact is that much of robust control theory is based on prior descriptions of model
uncertainty or model errors which classical identication theory has been incapable of delivering. Conversely, until recently identication theorists have
not spent much eort in trying to produce the
* Tel.: +32-10-47-25-90; fax: +32-10-47-21-80.
E-mail address: gevers@csam.ucl.ac.be (M. Gevers).
1
Our extensive quote should of course not be construed as
approval of the ideas expressed in that paper.

accurate uncertainty bounds around their estimated models that their robust control design colleagues were taking for granted. It is as if, until a few
years ago, the control design community and the
identication community had not been talking much
to each other. The gap between the surrealistic premises on which much of robust control design theory
is built and the failure of identication theory to
deliver accurate uncertainty bounds in the face of
unmodelled dynamics has brought to light major
deciencies in both theories, and a sudden awareness from around 1988 of the need to understand
better the interactions between both theories.
Surely, a natural place to search for an understanding of the interactions between identication
and robust control design is in the adaptive control
community. Indeed, adaptive control combines the
design of an on-line identier with that of a control
law . . . . . . An essential feature of adaptive control,
however, is that the identication is performed in
closed loop and that the controller therefore
impacts on the estimated model and on its quality
(i.e. its error with respect to the true system). It is
therefore to be expected that the separate designs of
the identier and of the controller without regard
for the eect of the control law on the identied

0959-1524/02/$ - see front matter # 2002 Elsevier Science Ltd. All rights reserved.
PII: S0959-1524(01)00018-X

520

M. Gevers / Journal of Process Control 12 (2002) 519531

model, or of the identied model on the robustness

of the control law, may not lead to a maximization
of the global robustness of the identier/controller
schema.
This last sentence was to become the program for
much of the research activity in the nineties: going from
separate designs to a synergistic design. In 1990, any
observer of the scene was aware of the many dierent
technical inconsistencies between the newly emerged
robust control theory and the more classical prediction
error identication theory. For example, prediction error
identication theory had very little to oer in terms of
explicit quantication of the model error, and whatever
tools that were available were incompatible with the
frequency domain uncertainty descriptions required for
robust control analysis and design.
However, the most crucial manifestation of the
identication/control gap was not so much these
technical incompatibilities, but rather the total absence
of synergy between the two parts of the design: identication design and model-based control design. The prevailing philosophy at the time was: First identify a
model with a method that also allows the estimation of
error bounds on this model; then design a controller
based on this model and its error bounds. The problem
is that an identication method whose sole merit is to
deliver error bounds on a restricted complexity model
may well produce a nominal model and an uncertainty
set that are ill-suited for robust control design.
Due to a lack of understanding of the interplay
between identication and robust control, most of the
earlier work focused on producing suitable estimates of
model quality (or uncertainty), and on bridging the gap
between identication and robust control. The most
obvious manifestation of this gap, and the one that triggered most of the research activity in the early nineties,
was the realization that robust control theory requires a
priori hard bounds on the model error, whereas classical
identication theory delivers at best soft bounds, i.e.
condence ellipsoids in a probabilistic sense. This led to
the development of new identication theories that were
called control-oriented only because they delivered
model uncertainty descriptions that were compatible with
those required by robust control theory. The question of
whether the identied models and their uncertainty
descriptions were likely to deliver high performance controllers was not addressed, at least initially.
It later became clear that the great hard-versus-soft
debate was not the real issue. To quote from another
plenary lecture [2]:
An identication and control design method that
leads to a closed loop system that is stable with
probability 99% is of course just as acceptable as
an H1-based design that leads to a guaranteed

stable closed loop, but that is based on prior error

bounds that cannot be veried.
However, even though the hard-versus-soft question
proved to be a non-issue, numerous other technical
stumbling blocks had to be conquered before robust
control analysis and design could be applied to models
identied from data, rather than just to models and
model uncertainty sets obtained from prior assumptions.
To summarize, the intense research eort of the nineties on identication for control has pursued two major
objectives:
. Obtaining better estimation procedures for the
quantication of the model uncertainty for identied
models; in particular, produce uncertainty descriptions that are compatible with robust control theory.
. Understanding the interaction between identication and model-based control in order to produce
control-oriented identication design guidelines.

In this paper, we shall mainly focus on the progress

accomplished in identication for control design, i.e. the
second issue. This line of research has led to such important new concepts as iterative model-based controller
redesign, cautious model and controller updates, and
eventually iterative model-free controller redesign. But,
before we venture in this direction, let us rst briey elaborate on the question of model uncertainty estimation,
if only to clearly distinguish it from the question of
identication for control.
The quantication of the model error is of course a
very important objective, whatever the goal of the identication step that has produced this model. A reputable
engineer should never deliver a product to his client
without some statement about the quality of that product, whether it be a machine tool, a measurement
device, or a dynamical model. When the product is a
model, and when the client is a robust control designer,
then this client expects a model quality statement that is
compatible with his/her robust control design tools.
There is no sense telling the robust control engineer that
the bias error on the delivered model can be implicitly
described by some complicated integral formula, and
that the noise-induced error is characterized by ellipsoidal condence regions on a meaningless parameter vector, if all the control engineer can handle for his robust
control design is a frequency domain error bound.
When that happens and this is exactly what did
happen ten years ago then the robust control engineer leaves the room in disgust and starts developing a
new identication theory which he calls control-oriented, only because it can deliver model error bounds that
are compatible with existing robust control theory.
All through the nineties, we have witnessed a tremendous activity, on the part of both communities, in the
area of model quality estimation and model uncertainty

M. Gevers / Journal of Process Control 12 (2002) 519531

description, with a view of bridging the technical

incompatibilities between the two theories. We cannot
possibly hope to reference the hundreds of relevant
papers. One of the better surveys of this line of research,
up to the middle of the decade, can be found in [3].
The results on model quality are necessary for the
construction of a synergistic design of identication and
robust control, but they do not constitute this synergy.
They are the technical building blocks. Indeed, identication for control is a design problem, as its name indicates. We now explain how to give meaning to this
problem, and why the solution leads to iterative controller
design.

2. From identication for control to iterative

identication and control
2.1. The setup
All through this paper we consider the situation where
there is a true system which, for the sake of analysis,
we assume to be linear time-invariant. For the sake of
simplicity, we also consider a single-inputsingle-output
system only in this paper. Thus, the true system is
represented by
S : yt G0 zut vt ;

where G0(z) is a linear time-invariant causal operator, y

is the measured output, u is the control input, and v is
noise, assumed to be quasistationary.
We now consider the situation where we can perform
experiments on this system with the purpose of designing
a feedback controller. We also consider that, most often,
the system is already under feedback control, and that
the task is to replace the present controller by one that
achieves better performance. This situation is representative of very many practical industrial situations.
It is also typical of many industrial applications that
the system to be controlled is very complex and possibly
nonlinear, and that it would, therefore, require a complex dynamical model to represent the system with high
delity. Any model-based control design procedure
would, therefore, lead to a complex or high order controller, since the complexity of a model-based controller
is of the same order as that of the system. The practical
situation, considered in this paper, is where we want the
to-be-designed controller to be linear and of low order.
2.2. In search of a low complexity controller
There are many ways of obtaining a low-complexity
controller for a high order system. These include identication, model reduction, or controller reduction, in
open- or in closed-loop, etc. A comparison between

521

identication methods and reduction methods, on an

industrial example, can be found in [4].
Here we consider the strategy which consists of identifying a low order model from data collected on the
real system, from which a model-based controller is
then computed. Given that the low order model cannot
possibly represent the true system over the whole frequency range, it will have a systematic error called the
bias error, in addition to the inevitable noise-induced
error called the variance error. This bias error and
indeed the total error must be taken into account in
the control design; hence the importance of producing
methods for the estimation of model errors. But what is
even much more important than estimating the model
error a posteriori is to design the identication in such a
way that the bias error does not harm the performance of
the controller that will be designed on the basis of this
approximate model. This is based on the observation that
one can design a high performance controller with a model
that has large error with respect to the real system (i.e. a
very wrong model), as long as this model represents with
high accuracy the dynamical features of the true system
that are essential for control design. For example, it is
essential that the model be very accurate around the
crossover frequency of the to-be-designed closed loop
system, but the error in its steady state gain can be huge.
The idea of tuning the bias error for control design is at
the core of identication for control. The practical formulation of the problem is one of nding an identication
criterion that takes account of the control objective. It is
an identication design problem, whose objective is to
produce, within a specied class of restricted complexity
models, a nominal model whose bias error distribution is
tuned towards the control design objective. As we shall
see later in this paper, this can only be achieved through
a succession of model and controller iterations; hence
the iterative schemes that have emerged in identication
for control.
The tuning of the bias error has led to iterative
schemes for the estimation of a control-oriented nominal model. However, to fully take advantage of robust
control theory, one must develop an identication for
control theory not just for the nominal model, but also
for the uncertainty regions around this nominal model.
Indeed, robust control is a model-based design methodology in which the controller is designed on the basis
of a nominal model together with an uncertainty region
around the nominal model: see e.g. [5]. It is the task of
model validation to construct an uncertainty region
around a nominal model. When the model and its
uncertainty region are to be used for robust control
design, then this validation step must also take account
of the control objective in such a way that the shape of
the uncertainty set also be tuned towards that control
objective. This is a much harder problem for which few
results are presently available. One exception is [6]

522

M. Gevers / Journal of Process Control 12 (2002) 519531

where an iterative scheme of H1 identication and

model-based control design is presented, in which an
uncertainty structure is chosen that enables one to
monitor robust performance. In the prediction error
framework, recent results have been obtained in [710].
We shall not elaborate on them in this paper, where we
focus on iterative designs.
2.3. Matching identication and control objectives
We have described the context in which we operate, and
we have introduced the motivation for identication for
control. We now show that the matching of identication
and control objectives leads to iterative identication and
control design. To illustrate the need for iterative design,
we take the simplest possible control design objective:
model reference control. Thus, consider the true system
(1)
and suppose we have identied a model G^ z
4

G z; ^ of G
 0, from some parametrized set of low
order models Gz; j 2 D , where D is a subset of
the Euclidean space. Consider a control law
ut Czrt
yt ;

and assume that our control design objective is to

design C(z) such that the closed loop transfer function
from vt to yt is some prespecied S(z). Then, given a
model G^ (z), the controller C(z) is computed from2
1
Sz:
^
1 GzCz

Compare the real closed-loop system of Fig. 1 with

the designed closed-loop system of Fig. 2, with both
loops driven by the same reference signal rt.
Now, looking at Figs. 1 and 2, one observes that:
G0 C
1
G^ C
rt ;
rt
vt ; y^ t
1 G0 C
1 G0 C
1 G^ C
C
C
rt

vt :
ut
1 G0 C
1 G0 C

Fig. 2. Designed (or nominal) closed-loop system.

The control performance error,3 dened as the error

between the actual and the designed outputs, is given
by:
"
#
G0 C
G^ C
1

vt
5
yt
y^ t
rt
^
1 G0 C 1 G C
1 G0 C
After some straightforward manipulations, this can be
rewritten as

 i
h
yt
y^ t Sz yt
G z; ^ ut :
6
Eq. (6) can be seen as an equality between a control
performance error on the left hand side (LHS) and a
ltered identication error on the right hand side
(RHS). Indeed, the RHS is a ltered (by S(z)) version of
the output error yt
G(z,^ )ut, where ut and yt are collected on the actual closed loop system of Fig. 1. Thus,
it appears that if  is obtained by minimizing the mean
square of the RHS of (6), i.e. by closed-loop identication with a lter S(z), then this will minimize the mean
square control performance error. In other words,
apparently (6) shows that we get a perfect match
between control error and identication error. However, life is more subtle and complicated. Indeed, the
controller C(z) is also a function of the model parameter
vector  via (3). Since the data collected on the real
closed-loop system of Fig. 1 are a function of C(z), they
are also dependent on . Hence, a more suggestive and
correct way to write (6) is as follows:
yt
y^ t Szyt 
Gz; ut 

yt

4
Even though the RHS of (7) looks like a closed-loop
prediction error, it cannot be minimized by standard
identication techniques, because  appears everywhere
and not just in G(z,).
We have illustrated the fact that with the simplest
possible control design mechanism, namely model
reference control, one can apparently equate a control
performance error to an identication error, but that
this identication error cannot be minimized by standard identication techniques because the parameter
vector appears in more than just the model G(z,). In

Fig. 1. Actual closed-loop system.

We assume for simplicity here that a causal solution exists for

C(z), since this is not the focal point of our discussion.

3
It was called that way in [11,12] because, if the closed-loop
transfer function of the actual system was equal to the reference model
S(z), this error would contain only the noise contribution.

M. Gevers / Journal of Process Control 12 (2002) 519531

other words, we know that, to make the control error

small, we should minimize the RHS of (7) with respect
to , but we dont know how to do this.
For optimization-based control design criteria, the
control performance criterion also denes an identication criterion that one would want to minimize with
respect to the model parameters: see [13,14]. This identication error is typically a norm of the following error:
yt
y^ t Sz; yt 
Gz; ut ;

where the data lter S(z,) is proportional to the sensitivity function of the design loop [compare with (3)] and
is now also -dependent. Again, we do not know how to
minimize the RHS of (8) with respect to .
As a consequence, the approach suggested in all
known identication for control schemes is to perform
identication and control design steps in an iterative
way, whereby the i-th identication step is performed on
ltered closed-loop data collected on the actual closedloop system with the (i
1)-th controller operating in the
loop. This corresponds to an i-th closed loop identication step in which the following ltered prediction error
is minimized with respect to , for xed ^ i
1 :


h




i
yt
y^ t S z; ^ i
1 yt ^ i
1
Gz; ut ^ i
1 :
9
We refer the reader to [2,14,15] for details and for a
survey on such iterative schemes.
Assume that one has chosen a model-based certainty
equivalence control design criterion, such that any
model G^ is mapped into a corresponding controller
C(G^ ), e.g. the Model Reference criterion above. Then an
interesting question is whether these iterative identication and control schemes converge
to the minimum of
4
the achieved cost over the set C
CGz; 8 2 D of
all such certainty equivalence controllers. This corresponds to asking whether by iteratively minimizing over
 the mean square of the prediction errors dened by
(9), one will converge to the minimum of
2
4 
J E Sz; yt 
Gz; ut  :

10

This question was analyzed in [16] for model reference

control; it was shown there that the iterative identication and control schemes do not generically converge to
the minimum of the achieved control cost.
This does not mean that iterative identication and
control schemes need to be thrown out of the window. In
fact, the idea of using available data, collected on the
actual closed loop system, to obtain a model that is better
suited for the design of a new controller, has found
immediate and widespread applications because it is easy
and intuitively reasonable. In typical applications large
numbers of closed loop data are owing into the control

523

computer, and it makes a lot of sense to use these data

to replace the existing controller by one that achieves
better performance. In addition, the theoretical work on
identication for control has shown that, in order to
compute a new controller with better performance, the
optimal experiment is to perform closed loop identication. Thus, no special experiments are required, and
there is no need to open the loop in order to design
the new model and the corresponding new controller.
Thus, this is one area where the transfer of technology
from theoretical research to applications has been
extremely fast. The rst applications of control-oriented
identication and iterative model-based controller tuning were reported within months after the theoretical
results had been produced. Representative examples can
be found in [1722]. The practical impact of iterative
closed loop identication and controller redesign has
been assessed in [23], where some interesting observations are made on the distinction between this batch-like
mode of operation and the more classical theory and
methods of adaptive control.
The guidelines that emerged during the nineties for
the application of iterative identication and control
schemes were supported by intense research that
brought to light two essential features.
. The benets of closed loop identication when the
model is identied with a view of designing a new
controller that is based on both the inputoutput
and the noise model: see e.g. [2430]. This produced a revival of interest in the design of closed
loop identication methods: see e.g. [3134].
. The need for cautious adjustments between successive
model and controller updates, in order to guarantee
closed loop stability or performance improvement
with the new controller: see e.g. [3538].

Despite its practical successes, the failure of all attempts

to establish convergence of iterative identication and
control schemes was a major worry, more from a theoretical than from a practical point of view. Indeed, in practice it was observed that major improvements in
performance of the closed-loop systems were obtained
within the rst few iterations, after which the improvements were very minor. Divergence typically occurred
only if one continued to iterate beyond these initial steps.
It is the analysis of [16] that revealed the reason for
the possible divergence. This analysis led the authors to
reformulate the iterative identication and control
design scheme as a parameter optimization problem, in
which the optimization is carried directly on the controller parameters, thereby abandoning the identication step altogether. This idea led to a gradient-based
algorithm for the iterative optimization of the parameters of any restricted complexity controller [39], which
was later called IFT, for iterative feedback tuning. In the

524

M. Gevers / Journal of Process Control 12 (2002) 519531

next section we describe the IFT algorithm and some of

its more recent developments.

3. IFT: a model-free iterative controller tuning method

3.1. Introduction to the IFT algorithm
The key feature of the IFT algorithm is that an
unbiased estimate of the gradient of a control
performance criterion is computed from signals obtained
from closed-loop experiments with the present controller
operating on the actual system. For a controller of given
(typically low-order) structure, the minimization of the
criterion is then performed iteratively by a GaussNewton
based scheme. For a two-degree-of-freedom controller,
three batch experiments are performed at each step of the
iterative design. The rst and third simply consist of
collecting data under normal operating conditions; the
only real experiment is the second batch which requires
feeding back, at the reference input, the output measured during normal operation. Hence the acronym
iterative feedback tuning (IFT) given to this scheme. No
identication procedure is involved.
The optimal IFT scheme, whose key idea is due to
Hjalmarsson, was initially presented in [39]. Given its
simplicity, it became clear that this new scheme had
wide-ranging potential, from the optimal tuning of simple PID controllers to the systematic design of controllers of increasing complexity that have to meet some
prespecied specications. In particular, the IFT method
is appealing to process control engineers because the
controller parameters can be successively improved
without ever opening the loop. In addition, in many
process control applications the objective of the controller design is to achieve disturbance rejection. With
the IFT scheme the tuning of the controller parameters
for disturbance rejection is driven by the disturbances
themselves; there is no need for the injection of a persistently exciting reference signal as is typically the case
in closed loop identication.
Since 1994, much experience has been gained with the
IFT scheme:
. It has been shown to compare favourably with
identication-based schemes in simulation examples: see [39], and its accuracy has been analyzed in
[40].
. It has been successfully applied to the exible
transmission benchmark problem posed by I.D.
Landau for ECC95, where it achieved the performance specications with the simplest controller
structure [41].
. It has been applied to the control of a exible arm of
the Laboratoire dAutomatique de Grenoble [42], on
a ball-on-beam system [43], for the temperature

control of a water tube and for the control of a

suspended plate [44], for the controller tuning in
cold rolling mills [45], for the tuning of a thermal
cycling module [46] and many more.
. It has been adapted to linear time-invariant
MIMO systems [47] and to time varying, and in
particular periodically time-varying, systems [48].
. It has been applied by the chemical multinational
Solvay S.A. to the tuning of PID controllers for a
number of critical control loops for which opening
the loop or creating limit cycles for PID tuning
was not allowed [49].
Here we present the fundamentals of the IFT algorithm, and we then review the performance achieved by
the scheme at S.A. Solvay.
3.2. The basic control design criterion
We present here a basic version of the IFT algorithm;
we refer to [50] for a more complete derivation and discussion. We consider the unknown true system (1), to be
controlled by a two degrees of freedom controller:
ut  Cr rt
Cy yt 

11

where Cr() and Cy() are linear time-invariant transfer

functions parametrized by some parameter vector 
2 Rn , and {rt} is an external reference signal, independent of {vt}: see Fig. 3. We use the notation yt(), ut()
for the output and the control input of the system (1) in
feedback with the controller (11), in order to make
explicit the dependence of these signals on the controller
parameter vector .
Let ydt be a desired output response to a reference
signal rt for the closed-loop system. This response may
be dened as the output of a reference model Td, i.e. ydt
=Tdrt, but for the IFT method knowledge of the signal
ydt is sucient. The error between the achieved and the
desired response is

Cr G0
4
d
d
r t
yt
y~ t  yt 
yt
1 Cy G0
1
vt :
12

1 Cy G0
If a reference model is used this error can also be
written as

Fig. 3. Block diagram of the closed-loop system.

525

M. Gevers / Journal of Process Control 12 (2002) 519531

y~ t 

Cr G0
1 Cy G0

Td rt

1
vt :
1 Cy G0

"

13

This error consists of a contribution due to incorrect

tracking of the reference signal rt and an error due to
the disturbance. With IFT the following quadratic control performance criterion is used:
"
#
N
N
X
X
1
2
2
E
y~ t  l ut 
J
14
2N
t1
t1
but any other dierentiable signal-based criterion can be
used. In (14) E[.] denotes expectation w.r.t. the weakly
stationary disturbance vt. The optimal controller parameter  is dened by
 arg minJ;

15

The errors y~ t  and ut() appearing in the criterion

can be ltered by frequency weighting lters Ly and Lu
to give added exibility to the design: see [50].
Let T0() and S0() denote the achieved closed loop
response and sensitivity function with the controller
{Cr(), Cy()}, i.e.
T0 

Cr G0
;
1 Cy G0

S0 

1
:
1 Cy G0

16

Given the independence of rt and vt, J() can be written

as
J

N 
2 1 h
2 i
1 X
ydt
T0 rt E S0 vt
2N t1
2
"
#
N
X
1
E
l
ut 2 :
2N
t1

The rst term is the tracking error, the second term is

the variance of the disturbance contribution, and the
last term is the penalty on the control eort. Observe
that the rst term contains no randomness, while the
second term is the variance of a stationary process,
which therefore needs no summation.
3.3. Criterion minimization
We now address the minimization of J() given by
(14) with respect to the controller parameter vector 
for a controller of specied structure. It is evident from
(12) that J() depends in a fairly complicated way on ,
on the unknown system G0 and on the unknown spectrum of {vt}. To obtain the minimum of J() we would
like to nd a solution for  to the equation

N
N
X
X
@J
1
@y~ t
@ut
 E
 l ut 
 :
y~ t 
@
N
@
@
t1
t1

17
@J
@

If the gradient
could be computed, then the solution of (17) would be obtained by the following iterative
algorithm:
i1 i
i R
1
i

@J
i :
@

18

Here Ri is some appropriate positive denite matrix,

typically a GaussNewton approximation of the Hessian
of J, while the sequence  i must obey some constraints
for the algorithm to converge to a local minimum of the
cost function J(): see [39].
Such problem can be solved by using a stochastic
approximation algorithm of the form (18) such as sug@J
gested in [51], provided the gradient @
(i) evaluated at
the current controller can be replaced by an unbiased
estimate. In order to solve this problem, one thus needs
to generate the following quantities:4
1. the signals y~ i and u(i);
@y~
@u
2. the gradients @
(i) and @
(i);
~
3. unbiased estimates of the products y~ i @y
@(i) and
@u
u(i) @
(i).
The computation of the last two quantities has always
been the key stumbling block in solving this direct optimal controller parameter tuning problem. The main
contribution of [39] was to show that these quantities
can indeed be obtained by performing experiments on
the closed loop system formed by the actual system in
feedback with the controller {Cr(i), Cy(i)}. This is
done as follows.
3.4. The IFT algorithm
At iteration i of the controller
 tuning algorithm, the
4
controller Ci
Cr i ; Cy i operates on the actual
plant. We then perform three experiments, each of
which consists of collecting a sequence of N data. Two
of these experiments (the rst and third) just consist of
collecting data under normal operating conditions; the
second is a real (i.e. special) experiment.
We denote N
length reference signals by rji , j=1, 2, 3, and the corresponding output signals by {yj(i)}, j=1, 2, 3. Thus we
have
r1i r;
r2i r
y1 i ;
r3i r;
4

y1 i T0 i r S0 i v 1i
y2 i T0 i r
y1 i
S0 i v2i
yielding y3 i T0 i r S0 i v3i :
yielding
yielding

To ease up the rather heavy notations, we have dropped the subscript t from all time signals in the following expressions.

526

M. Gevers / Journal of Process Control 12 (2002) 519531

Here vji denotes the disturbance acting on the system

during experiment j at iteration i. These experiments
yield an exact realization of y~ i :
y~ i y1 i
yd ;

19

while it is shown in [50] that

b
@u
1
4
i
@
Cr i



@Cy
@Cy
@Cr
3
2
i y i

i

i y i
@
@
@

20

@y
is an unbiased estimate of @
(i).
The three experiments described above generate corresponding control signals:


u1 i S0 i Cr i r
Cy i v1i ;





u2 i S0 i Cr i r
y1 i
Cy i v2i ;


u3 i S0 i Cr i r
Cy i v3i :
These signals can similarly be used to generate the
estimates of the input related signals required for the
estimation of the gradient (17). Indeed, u1(i) is a perfect
realization of u(i),

As a result, the controller parameters converge under

reasonable conditions to a stationary point of the performance criterion, provided the sequence of controllers
along the way are all stabilizing: see [50].
A number of implementation issues as well as design
choices are addressed in detail in [50]. They concern the
choice of the step size gi and of the matrix Ri in (24), the
choice of frequency weighting lters, the elimination of
possibly unstable controllers in the ltering operations
(20) and (22), the enforcement of integral action, the
attenuation of the eect of disturbances, as well as the
simplication that occurs in the case of a one-degree-offreedom controller, where the third experiment is not
necessary. One interesting design parameter is the step
size, which determines how much the controller changes
from one iteration to the next one. Before implementing
a new controller one can compare its Bode plot with
that of the previous one, and possibly reduce the step
size if one feels that the change is too large.

4. Applications of IFT in chemical process control

ui u1 i :

21

while
b
@u
1
4
i
@
Cr i



@Cy
@Cy
@Cr
3
2

i

i u i
i u i
@
@
@

22

@u
is an unbiased estimate of @
(i). An experimentally based
estimate of the gradient of J can be formed by taking
!
N
b
ct
ct
@J
@y
@u
1X
i
y~ t i
i lut i
i :
23
@
N t1
@
@

The next controller parameters are then computed by

replacing, in the iteration (18), the gradient of the cost
criterion by this estimate:
i1 i
i R
1
i

positive denite matrices. The key feature of our con@b

@J
J
(i), and also the motivation for the third
struction of @
experiment, is that this estimate is unbiased:
"
#
b
@J
@J
i i ;
25
E
@
@

b
@J
i
@

24

where { i} is a sequence of positive real numbers that

determines the step size, and where {Ri} is a sequence of

The IFT scheme has been applied by the chemical

multinational Solvay S.A. for the optimal tuning of
industrial PID controllers operating on a range of different control loops. In each of these loops, PID controllers were already operating. Important performance
improvements were achieved using the IFT method,
both in tracking and in regulation applications. The
reductions in variance achieved after a few (typically 2
to 6) iterations of the algorithm ranged from 25% in a
ow regulation problem in an evaporator, to 87% in a
temperature control problem for the tray of a distillation column. Here we present the results obtained with a
temperature regulation problem for a tray of a distillation column. An application to a ow control problem
in an evaporator is presented in [50].
The controller used was an industrial 2-degree-offreedom PID controller where the derivative action is
applied to y only, and where a rst order lter is applied
to y in order to limit the gain of the controller at high
frequencies when the derivative action is used. The time
constant of this lter is chosen as Td/8, Td being the
derivative time constant. The PID regulator parameters
were iteratively tuned using the IFT scheme, with the
following design choices: Gauss-Newton direction, stepsize  i=1 8i , control weighting l=0, sampling period of

527

M. Gevers / Journal of Process Control 12 (2002) 519531

Fig. 4. Control error over a 24-h period before tuning and after 6 iterations of IFT.

Fig. 5. Histogram of control error over 2-week period before tuning and after 6 iterations of IFT.

15 s, rd=yd=0 during 5 h. The deadtime and the time

constants of the process were unknown.
Fig. 4 presents temperature deviations with respect to
setpoint in a tray of a distillation column, over a 24-h
period, rst with the original PID parameters, then with
the PID controller obtained after 6 iterations of the new
scheme. Fig. 5 shows the corresponding histograms of
these deviations over 2-week periods. The control error
has been reduced by 70%.
Fig. 6 shows the Bode plots of the two-degree of
freedom controller (Cr, Cy) before optimal tuning (full
line), after 3 iterations of the IFT algorithm (dashed
line) and after 6 iterations (dotted line). The gain was
too low and the derivative action underused. Observe
that both the feedback controller Cy and the feedforward controller Cr are adjusted as a result of the IFT
iterations.
Table 1 shows the measured cost J with the 6 successive controllers, as well as the predicted value of the
cost, calculated at each iteration with the new controller
parameters, as explained above. The prediction was
good except for the 2nd iteration which was perturbed
by an abnormal disturbance.

5. Minimizing the settling time with IFT

The criterion (14) is well suited when the objective is
to follow a specic reference trajectory, but is not so

Table 1
Calculated and predicted cost
Iteration

Cost (measured)

Next cost (predicted)

1
2
3
4
5
6

0.80
1.00
0.57
0.37
0.22
0.14

0.36
0.59
0.35
0.18
0.15
0.11

appropriate if the objective is to change the output from

one setpoint to another one. Indeed, in such case the
goal is typically to reach the new setpoint with a minimum settling time, and one does not care about the
transient trajectory, provided it does not produce too
much overshoot. By constraining the output to follow
some particular reference trajectory yd during the transient, one puts too much emphasis on the transient
phase of the response at the expense of the settling time
at the new setpoint value.
To cope with this situation Lequin observed in [49]
that one can add non-negative weighting factors to each
element of y~ t and ut in the criterion (14). A simple way
to obtain a satisfactory closed loop response to a
desired setpoint change is then to set the weighting factors on y~ t to zero during the transient period and to one
afterwards, while the weights on the control are put to
one everywhere:

528

M. Gevers / Journal of Process Control 12 (2002) 519531

Fig. 6. Bode diagram of the 2-degree-of-freedom controller before tuning (full), after 3 iterations (dashed) and after 6 iterations of the algorithm
(dotted).

"
#
N
N
X
X
1
2
2
E
Jm 
y~ t  l ut  :
2N
tt0
t1
We say in such a case that a mask of length t0 is put
on the transient response of the tracking error. Often it
is not known a priori how much time is required to
achieve a setpoint change without overshoot. In such
case, one can perform the IFT iterations by initially
applying a long mask, and then gradually reducing the
length of this mask until oscillations start occurring.
We illustrate this idea with an example presented in
[52]. Consider the plant
1
Gs 2
s 0:1s 1
One wishes to tune a PID controller in order to
achieve a settling time of 20 s for the closed loop system.
The initial PID parameter values were taken as
K=0.025, Ti=2 and Td=1, yielding the very sluggish
response shown in Fig. 7.
The classical IFT criterion was then applied with a
desired response shown in dotted line in Fig. 8, with the
achieved response shown in full line on that same gure.
This response is very unsatisfactory, due to an unfortunate
choice of initial parameters.

The IFT criterion was then applied with a mask of

decreasing length, with an initial length of 80 s, and with
the same initial parameters. At every iteration of the
IFT scheme, the length of the mask was decreased by 20
s, until a mask of length 20 was reached. This led to the
closed-loop response shown in Fig. 9.
Observe the dramatic improvement of the closed loop
response.

6. Conclusions
Iterative redesign of controllers using data collected
on the operating closed loop system has emerged as a
new, powerful and successful control design methodology, as a result of signicant progress accomplished in
the nineties on the understanding of the interplay
between identication and control design. Most of the
schemes are based on model and controller updates; they
require safeguards such as cautious changes between
successive controllers. The study of these identication
for control schemes has somewhat surprisingly also led
to iterative schemes that are entirely model-free.
In this paper we have focused on the design of the
nominal model and/or controller via these iterative
schemes, since these have given rise to the more practical

M. Gevers / Journal of Process Control 12 (2002) 519531

529

Fig. 7. Closed-loop step response with initial PID parameters.

Fig. 8. Optimal closed-loop step response (full) obtained with the classical IFT criterion and using the desired response (dashed).

Fig. 9. Optimal closed loop step response obtained with the IFT criterion using masks of decreasing length.

design methods, well suited for process control applications. We have barely touched upon the vast amount of
progress accomplished on model uncertainty estimation,
and have completely left aside our recent theoretical
work on model and controller validation.

Triest, P.M.J. Van den Hof, V. Wertz, and Z. Zang. The

author acknowledges the Belgian Programme on Interuniversity Poles of Attraction, initiated by the Belgian
State, Prime Ministers Oce for Science, Technology and
Culture. The scientic responsibility rests with the author.

Acknowledgements

References

The work reported herein is the result of collaborations with a large number of co-authors over the last 10
years. I would like to acknowledge B.D.O. Anderson, P.
Ansay, R.R. Bitmead, X. Bombois, B. Codrons, F. De
Bruyne, G.C. Goodwin, S. Gunnarsson, H. Hjalmarsson, C. Kulcsar, J. Leblond, O. Lequin, L. Ljung, B.
Ninness, A.G. Partanen, M.A. Poubelle, G. Scorletti, L.

[1] M. Gevers, Connecting identication and robust control: a new

challenge, in: Proceedings of the IFAC/IFORS Symposium on
Identication and System Parameter Estimation. Budapest,
Hungary, 1991, pp. 110.
[2] M. Gevers, Towards a joint design of identication and control?
in: H.L. Trentelman, J.C. Willems (Eds.), Essays on Control:
Perspectives in the Theory and its Applications, Birkhauser, New
York, 1993, pp. 111151.

530

M. Gevers / Journal of Process Control 12 (2002) 519531

[3] B. Ninness, G.C. Goodwin, Estimation of model quality, Automatica 31 (12) (1995) 3274.
[4] P. Bendotti, B. Codrons, C. M. Falinower, M. Gevers, Control
oriented low order modelling of a complex PWR plant: a comparison between open loop and closed loop methods, in: Proceedings of the 37th IEEE Conference on Decision and Control,
Tampa, Florida, 1998, pp. 3390-3395.
[5] K. Zhou, J.C. Doyle, K. Glover, Robust and Optimal Control,
Prentice Hall, New Jersey, 1995.
[6] R.A. de Callafon, P.M.J. Van den Hof, Suboptimal feedback
control by a scheme of iterative identication and control design,
Mathematical Modelling of Systems 3 (1) (1997) 77101.
[7] M. Gevers, X. Bombois, B. Codrons, F. De Bruyne, G. Scorletti,
The role of experimental conditions in model validation for control, in: A. Garulli, A. Tesi, A. Vicino (Eds.), Robustness in
Identication and Control Proceedings of Siena Workshop,
July 1998, Vol. 245 of Lecture Notes in Control and Information
Sciences, Springer, Berlin, 1999, pp. 7286.
[8] X. Bombois, M. Gevers, G. Scorletti, A measure of robust stability for an identied set of parametrized transfer functions, IEEE
Transactions on Automatic Control 45 (11) (2000) 21412145.
[9] X. Bombois, M. Gevers, G. Scorletti, B.D.O. Anderson,
Robustness analysis tools for an uncertainty set obtained by prediction error identication. Automatica, in press.
[10] M. Gevers, X. Bombois, B. Codrons, G. Scorletti, B.D.O.
Anderson, Model validation for control and controller validation: a prediction error identication approach, submitted for
publication to Automatica.
[11] K.J. Astrom, Matching criteria for control and identication, in:
Proceedings of the 2nd European Control Conference, Groningen, The Netherlands, 1993, pp. 248251.
[12] K.J. Astrom, J. Nilsson, Analysis of a scheme for iterated identication and control, in: Proceedings of the IFAC Symposium on
Identication, Copenhagen, Denmark, 1994, pp. 171176.
[13] Z. Zang, R.R. Bitmead, M. Gevers, Iterative weighted leastsquares identication and weighted LQG control design, Automatica 31 (11) (1995) 15771594.
[14] P.M.J. Van den Hof, R.J.P. Schrama, Identication and controlclosed-loop issues, Automatica 31 (1995) 17511770.
[15] R.R. Bitmead, Iterative control design approaches, in: Proceedings of the 12th IFAC World Congress, Sydney, Australia, 1993,
Vol. 9, pp. 381384.
[16] H. Hjalmarsson, S. Gunnarsson, M. Gevers, Optimality and suboptimality of iterative identication and control design schemes,
in: Proceedings of the American Control Conference 95, 1995b,
Vol. 4, pp. 25592563.
[17] A.G. Partanen, R.R. Bitmead, The application of an iterative
identication and controller design to a sugar cane crushing mill,
Automatica 31 (1995) 15471563.
[18] R.J.P. Schrama, O.H. Bosgra, Adaptive performance enhancement by iterative identication and control design, Int. Journal of
Adaptive Control and Signal Processing 7 (5) (1993) 475487.
[19] R.A. de Callafon, P.M.J. Van den Hof, M. Steinbuch, Controlrelevant identication of a compact disc pick-up mechanism, in:
Proceedings of the 1993 IEEE Conference on Decision and Control. San Antonio, Texas, 1993, pp. 20502055.
[20] R.A. de Callafon, Feedback Oriented Identication for Enhanced
and Robust Control: A Fractional Approach Applied to a Wafer
Stage. PhD thesis. Delft University of Technology, 1998.
[21] U. Holmberg, S. Valentinotti, D. Bonvin, An identication-forcontrol procedure with robust performance, Control Engineering
Practice 8 (2000) 11071117.
[22] B.L. Cooley, J.H. Lee, Control-relevant experiment design for
multivariable systems described by expansions in othonormal
bases, Automatica 37 (2) (2001) 273281.
[23] I.D. Landau, From robust control to adaptive control, Control
Engineering Practice 7 (9) (1999) 11131124.

[24] K. Liu, R.E. Skelton, Closed loop identication and iterative

controller design, in: 29th IEEE Conference on Decision and
Control, 1990, pp. 482487.
[25] R.J.P. Schrama, Accurate identication for control: The necessity
of an iterative scheme, IEEE Trans. on Automatic Control 37
(1992a) 991994.
[26] R.G. Hakvoort, R.J.P. Schrama, P.M.J. Van den Hof, Approximate
identication with closed-loop performance criterion and application to LQG feedback design, Automatica 30 (4) (1994) 679690.
[27] W.S. Lee, B.D.O. Anderson, I.M.Y. Mareels, R.L. Kosut, On
some key issues in the windsurfer approach to adaptive robust
control, Automatica 31 (11) (1995) 16191636.
[28] H. Hjalmarsson, M. Gevers, F. De Bruyne, For model-based
control design, closed-loop identication gives better performance, Automatica 32 (1996) 16591673.
[29] U. Forssell, L. Ljung, Some results on optimal experiment design,
Automatica 36 (2000) 749756.
[30] I.D. Landau, Identication in closed-loop: a powerful design tool
(better design models, simple controllers), Control Engineering
Practice 9 (1) (2001) 5165.
[31] F. Hansen, G. Franklin, R. Kosut, Closed-loop identication via
the fractional representation: Experiment design, Proc. ACC, 1989.
[32] P.M.J. Van den Hof, R.J.P. Schrama, An indirect method for
transfer function estimation from closed loop data, Automatica
29 (1993) 6, 15231527.
[33] P.M.J. Van den Hof, R.J.P. Schrama, R.A. de Callafon,
O.H. Bosgra, Identication of normalized coprime plant factors
from closed-loop experimental data, European Journal of Control 1 (1995) 6274.
[34] U. Forssell, L. Ljung, Closed-loop identication revisited, Automatica 35 (1999) 12151241.
[35] R.J.P. Schrama, Approximate identication and control design,
PhD thesis, 1992.
[36] R.R. Bitmead, M. Gevers, A.G. Partanen, Introducing caution in
iterative controller design, in: 11th IFAC Symposium on System
Identication, Fukuoka, Japan, 1997, pp. 1701-1706.
[37] B.D.O. Anderson, X. Bombois, M. Gevers, C. Kulcsar, Caution
in iterative modeling and control design, in: IFAC Workshop on
Adaptive Control and Signal Processing, Glasgow, 1998.
[38] R.R. Bitmead, B.D.O. Anderson, M. Gevers, L.C. Kammer,
Cautious controller tuning, submitted for publication.
[39] H. Hjalmarsson, S. Gunnarsson, M. Gevers, A convergent iterative restricted complexity control design scheme, in: Proceedings
of the 33rd IEEE Conference on Decision and Control, Orlando,
Florida, 1994, pp. 1735-1740.
[40] H. Hjalmarsson, M. Gevers, Frequency domain expressions of
the accuracy of a model-free control design scheme, in: Proceedings of the 11th IFAC Symposium on System Identication, Vol.
1, Fukuoka, Japan, 1997, pp. 135-140.
[41] H. Hjalmarsson, S. Gunnarsson, M. Gevers, Model-free tuning
of a robust regulator for a exible transmission system, European
Journal of Control 1 (2) (1995) 148156.
[42] B. Ceysens, B. Codrons, Synthe`se Iterative de Controleurs sans
Identication. Technical report, Masters thesis, Universite Catholique de Louvain, 1995.
[43] F. De Bruyne, P. Carrette, Synthetic generation of the gradient
for an iterative controller optimization method. CD-ROM of 4th
European Control Conference, paper THA-F2, 1997.
[44] D. Molenaar, Model-free Data-driven Optimal Tuning of Controller Parameters: A Practical Guide with Application Examples.
Technical report. Report LiTH-ISY-R-1722, Department of
Electrical Engineering. Linkoping University, Sweden, 1995.
[45] H. Hjalmarsson, M.T. Cameron, Iterative feedback tuning of controllers in cold rolling mills, in: Proceedings of the 14th IFAC World
Congress, paper I-3b-08-5. Beijing, PR China, 1999, pp. 445450.
[46] K. El-Awady, A. Hansson, B. Wahlberg, Application of iterative
feedback tuning to a thermal cycling module, in: Proceedings of

M. Gevers / Journal of Process Control 12 (2002) 519531

the 14th IFAC World Congress, paper I-3b-08-4. Beijing, PR
China, 1999, pp. 439444.
[47] H. Hjalmarsson, Ecient tuning of linear multivariable controllers using iterative feedback tuning, International Journal of
Adaptive Control 13 (1999) 553572.
[48] H. Hjalmarsson, Model-free tuning of controllers: experience
with time-varying linear systems. Proceedings of the 3rd European Control Conference, 1995, pp. 28692874.
[49] O. Lequin, Optimal closed loop PID tuning in the process industry with the Iterative Feedback Tuning scheme, in: CD-ROM of

531

European Control Conference, paper TH-A-H6. Brussels, Belgium, 1997.

[50] H. Hjalmarsson, M. Gevers, S. Gunnarsson, O. Lequin, Iterative
feedback tuning: theory and applications, IEEE Control Systems
Magazine 18 (1998) 2641.
[51] H. Robbins, S. Monro, A stochastic approximation method,
Ann. Math. Stat. 22 (1951) 400407.
[52] O. Lequin, M. Gevers, L. Triest, Optimizing the settling time with
Iterative Feedback Tuning, in: Proceedings of the 14th IFAC World
Congress, paper I-3b-08-3. Beijing, PR China, 1999, pp. 433437.